Drop 3/14: Our Conversational Agents platform, Samvaad, now powers over a million minutes of interactions daily. Every month our TAM estimates keep increasing as we discover new use cases, not thought possible earlier - population scale outreaches, hybrid onboarding journeys over phone and WhatsApp, 24*7 sales assistants, ... Here... is a real conversation between a merchant and B2B marketplace support agent.show more

Some key learnings from the journey: First, voice AI should be thought of as core infrastructure, which will power an increasing fraction of human-computer interaction. And this infrastructure needs to guarantee low latency, high reliability, and extreme scalability. Our experience scaling up to 45M interactions with an insurance company in a matter of a week strengthened our infrastructure game. We today serve our models on self-managed GPUs with consistent latency no matter the call load.

Second, human-likeness scores high marks in effectiveness of voice interactions. With a tight integration between our AI research and Samvaad product teams, we have unlocked levers which cause 80% of our calls to be indistinguishable from human callers. Indeed, our agents are providing over 5 percentage points higher engagement across customer services use cases, and almost doubling capturing of sales interest.

Third, beyond conversational agents being engaging for customers, they are truly proving to be insightful for companies. By analysing conversational logs at scale with the help of AI, statistically sound inferences are being derived. For instance, our work with a credit card company helps distill reasons why a user chooses not to buy a card. This truly is listening at scale, wherein the data dividends of a call far outlasts the conversation itself.

Fourth, we discovered the power of multiple agents working together towards a common goal. These agents could be spread across time and/or channels - mimicking multiple personas working through a choreographed outreach. Or, more interestingly, simultaneous interactions on multiple channels such as a voice call on speaker plus an interactive WhatsApp bot. This combines the power of voice to squash IVR hierarchies while leveraging text for precise inputs. This is already live with Tata Capital that is using this to onboard users.

Finally, we discovered some interesting scaling laws of conversational agents. Some agents, when well built, hit scale. That scale provides feedback on how to improve the agents further, which in turn helps them reach even greater scale. This reinforcement can be rapid. We have continued to discover these “rocketship” agents with increasing frequency.

What rocketship conversational agents will you build? Head over to our product page to learn more -

Building Samvaad and watching it scale exponentially has been a highlight of my life! I loved every minute of the experience and I wouldn’t trade it for the world. Super excited to see it cross 10M minutes a day soon. Huge shoutout to the entire Samvaad who have given their blood, sweat, and tears to make this happen.

Loved this. A few questions: 1) Does it cut off the conversation if it recognizes the call is drifting too far off-topic when talking to the end user? I was talking to the "/api/app-runtime/orgs/sarvam-internal.ai/workspaces/sarvam-internal-ai-9191fd" appointment agent, and the conversation went in a different direction. This might happen with customers too as end users can let the call go on indefinitely, burning Rs. per/min costs. Is this configurable from the customer dashboard? 2) Not sure if I'm correct but ig I heard background noise from the Cart Recovery Agent (like what you'd normally hear when talking to a support rep, with background sounds clearly audible). Was this a detail you added? If so, it was a great addition. If not, it could be a good idea.

This is simply outstanding!! The fluency, reasonability of responses are all so ‘real’ How many languages can this support? Prima facie this can also be used multi lingual conversations if real time translation can be supported..also useful for learning languages..

tried the appointment booking implementation. it has good gaurd rails, didn't allow me to book the hospital for coding camp or schedule an appointment at 3 am. it did however gave me booking confirmation for jan 5, so i guess more gaurd rails are required.

Woah! Wouldn't have guessed that's a conversational agent. Amazing 👏 Indian voices for Indian market. 🏅Ur team should feel proud of their achievements

Why is the Hindi language Urdufied? In pure Hindi, there is no word like "maaf." Please don't butcher our own language!

This is too good!

Wow

When is Sarvam AI launching on play store ?? Is there any plan ??

Intresting but where can I test and evaluate website playstore and app Store links missing

Amazing!

Great one yet heard in hindi The voice, response, delay, steps thinking reasoning, accent, all things were into the perfect place

wow! feels unreal

This is amazing. I wonder if we can get to a world where AI does not summarize or verify whatever it has heard from the human. Humans don't talk like that so I wonder if that feature is in the pipeline or if there is a technical challenge to solving that. I can think that if you don't verify probably you will interpret it wrong, but I mean what are the chances that AI hears the wrong thing? Because verifying after every sentence is annoying. Maybe if the accuracy of AI hearing the right thing is say 90%, you save 9 out of those 10 repeated conversations .

Who do you use for GPUs?

fascinating!

At 1M mins/day, 3 things matter next: evals on real calls, cost+latency per resolved intent, and handoff quality. What’s your p50 latency + % deflected vs routed to a human?

Congrats, sir, on this milestone & continuous growth.

@grok please explain to me like i am 10

Kudos to entire team ☺️. Keep inspiring.

India's Elevenlabs!

🚀

Just wow

That’s impressive! It’s exciting to see how Samvaad is evolving and unlocking new possibilities for engagement.

Crazy

Ws

Incredible! Great work guys 👏🏼

over a million minutes daily? samvaad crushing it lfg. what's next use case?

You guys should just build a CRM now

@SarvamAI 🚀🚀

Great💯💯